The Death of the Stubborn Developer

Written past midnight with Gene Kim texting "PUBLISH IT!!" from the other end, and the sleep deprivation stays on the page: Steve attempts a pun about Idan Gazit's beef with chop, apologizes ("This is like slapping both myself and you to keep me awake"), and restarts the paragraph without cleaning up the wreck. The fun part is watching reality argue back mid-publish. Steve scoffs that nobody has heard "so much as a fart" from Devin, the overfunded poster child for autonomous coding agents, in the eight months since its launch — and Devin goes GA the very next day, forcing an inline update that he cheerfully reads as a point for his side, since the relaunch quietly shrank "it does everything" down to batch refactoring. He also lets the opposition speak through a spokesman he plainly likes, Gazit of GitHub Next, and backs his own case with numbers from Sourcegraph customer studies: a floor of 30% productivity gains, "even when the engineers are grumbling and think they are slower with it."

The page flags this one as a prediction miss, and the miss is worth locating exactly. Chop was real; the clock was wrong. Steve guesses chat-first programming will reign three years, possibly ten, and calls himself "bearish" on agents taking over the higher-level work — within a few months he was chronicling the agent takeover himself. What reads best now is, oddly, his complaint about chop's toil: all that manual fetching of context in and integrating of answers out is a precise requirements document for the tools that soon buried it. The Optionality section holds up too, with its observation that LLMs help you least at what you already do well and most at everything you'd been putting off.